llm_bridge

Open source by Santander AI Lab. A tiny, vendor-neutral LLM client library — one interface for OpenAI, DeepSeek, Alibaba Qwen, AWS Bedrock and Google Gemini (or bring your own AI backend).

Part of Santander AI Open Source — open source AI projects from Banco Santander (santander.com).

A tiny, vendor-neutral wrapper for any LLM backend. One small interface, pluggable providers. Write your application against LLMClient once and switch between OpenAI, DeepSeek, Alibaba Qwen, AWS Bedrock, Google Gemini, a local server, or your own internal backend — without touching your code.

• Canonical interface + thin SDK adapters. One contract (LLMClient) with a small, modular adapter per official vendor SDK. The SDKs do the heavy lifting (retries, streaming, types); llm_bridge just normalises the calls.
• Zero required dependencies. The core (mock, callable) is pure standard library. Each vendor SDK is an optional extra you install on demand.
• Bring your own backend. Wrap any function with the callable provider — the library never needs to know which backend you use.
• One interface. chat(messages) and complete(prompt) returning a normalised LLMResponse (content + token usage + latency).

Installation

pip install llm-bridge                 # core only (no vendor SDKs)
pip install "llm-bridge[openai]"       # + OpenAI SDK
pip install "llm-bridge[deepseek]"     # + OpenAI SDK for DeepSeek
pip install "llm-bridge[qwen]"         # + OpenAI SDK for Qwen/DashScope
pip install "llm-bridge[aws]"          # + AWS Bedrock (boto3)
pip install "llm-bridge[google]"       # + Google Gemini (google-genai)
pip install "llm-bridge[all]"          # everything

Quickstart

from llm_bridge import create_llm

# Offline, no credentials — great for tests and demos.
llm = create_llm({"provider": "mock"})
print(llm.complete("Hello!").content)

# Switch provider by changing one dict — your code stays the same.
llm = create_llm({"provider": "openai", "model": "gpt-4o-mini"})       # needs [openai] + OPENAI_API_KEY
llm = create_llm({"provider": "deepseek", "model": "deepseek-v4-pro"})  # needs [deepseek] + DEEPSEEK_API_KEY
llm = create_llm({"provider": "qwen", "model": "qwen-plus"})            # needs [qwen] + DASHSCOPE_API_KEY
llm = create_llm({"provider": "bedrock", "model": "<bedrock-model-id>"}) # needs [aws] + AWS creds
llm = create_llm({"provider": "google", "model": "gemini-2.5-flash"})   # needs [google] + GOOGLE_API_KEY

resp = llm.chat([
    {"role": "system", "content": "You are concise."},
    {"role": "user", "content": "Name three primary colors."},
])
print(resp.content, resp.total_tokens)

Bring your own LLM

The callable provider wraps any function — the recommended way to plug in a proprietary or internal backend:

from llm_bridge import create_llm

def my_backend(messages, temperature=0.7, max_tokens=1024, **kwargs):
    # call your own SDK / gateway / local model here, return the text
    return "the model output"

llm = create_llm({"provider": "callable", "callable": my_backend})
print(llm.complete("Hi").content)

Providers

Provider	Name(s)	Dependency
Mock (offline)	mock	none
Bring your own	callable	none
OpenAI (and OpenAI-compatible)	openai	[openai]
DeepSeek	deepseek	[deepseek]
Alibaba Qwen	qwen	[qwen]
AWS Bedrock (Converse)	bedrock, aws	[aws]
Google Gemini	google, gemini	[google]

Credentials are read from environment variables (OPENAI_API_KEY, DEEPSEEK_API_KEY, DASHSCOPE_API_KEY, GOOGLE_API_KEY/GEMINI_API_KEY, standard AWS credential chain). Never hardcode secrets.

The openai provider also targets any OpenAI-compatible endpoint (vLLM, Ollama, Azure OpenAI, or an internal gateway): pass a base_url (or set OPENAI_BASE_URL). For local servers without auth, set a dummy OPENAI_API_KEY.

The deepseek provider uses DeepSeek's OpenAI-compatible API via the OpenAI SDK, defaults to https://api.deepseek.com, and accepts base_url or DEEPSEEK_BASE_URL for compatible endpoints.

The qwen provider uses Alibaba Model Studio/DashScope's OpenAI-compatible API via the OpenAI SDK. It defaults to https://dashscope-intl.aliyuncs.com/compatible-mode/v1 and accepts base_url or DASHSCOPE_BASE_URL for other regions or workspaces.

The interface

class LLMClient:
    @property
    def model(self) -> str: ...
    @property
    def provider(self) -> str: ...
    def chat(self, messages, *, temperature=0.7, max_tokens=1024, **kwargs) -> LLMResponse: ...
    def complete(self, prompt, *, system=None, temperature=0.7, max_tokens=1024, **kwargs) -> LLMResponse: ...

@dataclass
class LLMResponse:
    content: str
    model: str
    prompt_tokens: int
    completion_tokens: int
    finish_reason: str
    latency_ms: float
    raw: Any            # the provider's raw response
    # .total_tokens

Adding a provider

Implement LLMClient, expose build(config) -> LLMClient, and register it in llm_bridge.registry. See CONTRIBUTING.md.

Examples

See examples/: mock_example.py, callable_example.py, openai_example.py, deepseek_example.py, bedrock_example.py, qwen_example.py, bedrock_example.py, google_example.py.

Requirements

• Python 3.9+
• No required runtime dependencies for the core (mock, callable).
• Optional vendor SDKs are installed on demand via extras ([openai], [deepseek], [qwen], [aws], [google], [all]).

Contributing

Contributions are welcome! Please read CONTRIBUTING.md and our CODE_OF_CONDUCT.md. External contributors will be asked to sign the CLA (handled automatically by the CLA Assistant bot).

Security

Please report vulnerabilities privately — see SECURITY.md. Do not open a public issue for security reports.

Citation

If you use this software, please cite it using the metadata in CITATION.cff.

Disclaimer

This software is an open source project from the Santander AI Lab, provided "as is" under its license, without warranties or conditions of any kind. It is not an official Banco Santander product or service, carries no commitment of production support, and does not constitute financial, legal or professional advice.

"Santander" and its logo are registered trademarks of Banco Santander, S.A. The project license does not grant any right to use them beyond factual attribution.

If you believe you have found a security vulnerability, follow our security policy — do not open a public issue. You are responsible for assessing the suitability of this software for your use case and for keeping your own deployments up to date.

License

Apache License 2.0 — see LICENSE and NOTICE.